Its not that long since I posted the last article on AI.  But some things have changed during the last few month. And they went all, I can’t describe it otherwise...insane.

Its a bit of a rambling, sorry for that:

More and bigger data centres

Some AI companies plan to build huge data centres in scale of small cities.

Going big time nuclear.

To power this a huge investment is planed, to build nuclear reactors numbered by the dozens. (not sure if they mean full fledged reactors or using some modular smaller design)

Now lets just stop the hype train a minute: All this has to be financed and operated, not only now by investors, but later by the consumer….plus profit..and the longer it takes the higher the price. Not to mention that new estimates of the lifetime of the used cards in current AI operation is estimated by 5 years. So a lot of the initial investment has also to be reinvested.

Just to put it into perspective, this is not a natural scale of operation. Most AI companies still play the investors game: Get as much people attached to the product, so they can run the numbers and see what price / profitability they have in future. If you got a running profitable business and want to scale up you might make bold investments, but you have also a solid base to project your future monetary expense and gain.

The price of AI

We don’t know the price of AI. Currently its the “free lunch time”.  It might be that it won’t be a sustainable for a lot of individuals and small companies once these AI providers actually want to make profit. Here is a small estimate on some current available figures:

Open AI lost 11.5 billion last quarter, that’s 125 million per day. They currently announced they have 1 million business customers.  It cost around 30 bucks per month of writing That means they need 125 times the current user base at current price, OR each existing account has to pay a little less than 4000 bucks per month, if my calculation is right, to be even out without any profits.  

That is with current hardware and software. Now everyone with big money can estimate how much time in workforce ChatGPT actually saved them, and if its viable for their business. And this price does not include future investments (interests) in those nuclear reactors and data cities.

Access to AI...or: is the infrastructure needed ?

For those who can’t afford it, like me, it means to be cut off of this technology….but are we ?

Lets take a look at the technology which helped to foster the AI hype: AI image generators
5 years (yes that 5 years people pull out of their sleeves when discussing AI)  after the introduction of the technology,  is save to say that everyone who invested money in them can forget about getting their money back.

Local image generators run on modest systems (even on CPU) are becomming, more or less, easy to install, and often offer more options.  It might be slower to use them, but since it often offers more direct feedback than the web interfaces and are more easy to be integrated into someones pipeline, the time lost is less of an issue.

Could this also happen to openAI / meta and google ?

Currently good video and LLMs do not run locally on machines with modest hardware, so they are save as for now.
But two things to look for are advancements in affordeable hardware (like the nvidia spark) and better optimised models. 
I mean the current price for a spark is also around 4k, which runs smaller models, This might be just enough for smaller companies to put into their own server farm. Price might be amortised in less than a quarter of a year, and the luxury not to send data into the net might be an additional benefit. And this does not include other companies which also develop AI chips.
For me it sounds more sane to invest in those technologies, than investing that amount of money on infrastructure. It feels a bit like running a logistic company with cars. Insteead of inventing a truck or investing in people who could create one,  they scale up the amount of cars.  Thats what a "Manhattan sized data centre" for me sounds like.

Last but not least, Broad models are fine, but in the end I bet more on smaller specialised and optimised models. (Eg a cooking AI, a text summarisation AI, and a coding AI, a friend AI etc). Which are used on a specific problem. Its cool that ChatGPT can joke and code, but if I just need it to summarise a pdf, it does not help me. Maybe we even get a dumb “ÜberAI” which delegates tasks to specific smaller AIs but I’m no AI coder.
And btw even if a certain task can't be done locally, it might be more cost effective to just outsource that task.

Also we already saw that companies under financial pressure already build optimised systems (DeepSeek). IF commercial entities raise their prices the pressure to develop smaller, more affordable, models does also rises. And its still uncertain if a local AI needs to run at the speed of a online system. Maybe a 4 to 6 times slower AI is still good enough for most people. Like the Concorde: It might was the fastest commercial Airliner, but the majority of people didn’t need it and choose a slower but more affordable option.

With a minimum of 6 years (lowest estimate) to build a nuclear reactor, a lot can happen. If you commit to a large data centre in that timeframe and money, you also commit to a singular strategy which might become outdated.  

And as someone who still knows Yahoo directory, and the omnipresent Altavista, and know what happened to them once a more efficient and easy to use system like google simply eradicated the market, I don’t put a bet on any company becoming the new market leader in AI in longterm at current time. 

Yes I’m still indifferent to AI, and no I don’t ignore the copyright problem, but this now for courts to decide.  But I fear this won’t be judged in favour to the individual creator, but to large companies with huge assets.